User Interaction Model for Concierge Content Display
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current online concierge systems face inefficiencies in displaying content items to users, as different users and content items have varying effectiveness in encouraging interactions, leading to resource wastage and potential discouragement of future interactions.
Innovation Solution
The online concierge system employs a method where users are selected for holdout groups for different content items, withholding display of these content items to users in the holdout groups while displaying them to others. This allows the system to determine interaction rates with and without content item display, generating training data for a user interaction model to predict the effectiveness of content item display on user interactions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If content items are displayed to all users, then the system can potentially encourage more user interactions, but resources are wasted on users unlikely to interact and user experience may deteriorate
Solution Approach 1:
The patent applies local quality by customizing content item display strategies for different user segments. Instead of uniform display to all users, the system identifies specific user groups (e.g., high-propensity vs. low-propensity users) and applies different display policies to each group, optimizing resource allocation and interaction rates locally for each segment
Solution Approach 2:
The system changes the parameter of content item display probability based on user characteristics and historical data. By adjusting the likelihood of displaying content items to different users, the system optimizes the balance between encouraging interactions and avoiding resource wastage on unlikely candidates
2Loss of energy
If content items are displayed to users unlikely to interact, then resource allocation may be inefficient, but displaying no content items may reduce potential interaction opportunities
Solution Approach 1:
The patent implements partial action by displaying content items to only a portion of users rather than all users. Specifically, it identifies and targets users with higher predicted interaction propensity, applying content item display selectively to this subset while withholding from others, thus achieving efficient resource utilization without completely eliminating interaction opportunities
3Measurement precision
If the system tracks user interactions with content items to improve targeting, then prediction accuracy improves, but system complexity increases
Solution Approach 1:
The patent implements feedback mechanisms by tracking user interactions with content items and using this data to refine prediction models. The system continuously collects interaction data, updates user propensity estimates, and adjusts content item display strategies accordingly, creating a closed-loop system that improves prediction accuracy over time
Solution Approach 2:
The system employs self-service by automatically collecting interaction data, training prediction models, and optimizing content item allocation without requiring manual intervention. The machine learning infrastructure autonomously processes data, generates predictions, and adjusts display strategies, reducing operational complexity despite increased system capabilities
Data Source
AI summary
An online concierge system trains a user interaction model to predict a probability of a user performing an interaction after one or more content items are displayed to the user. This provides a measure of an effect of displaying content items to the user on the user performing one or more interactions. The user interaction model is trained from displaying content items to certain users of the online concierge system and withholding display of the content items to other users of the online concierge system. To train the user interaction model, the user interaction model is applied to labeled examples identifying a user and value based on interactions the user performed after one or more content items were displayed to the user and interactions the user performed when one or more content items were not used.